package/x265/0002-arm-asm-primitives.patch

165          p.cu[BLOCK_4x4].copy_cnt     = PFX(copy_cnt_4_neon);
167          p.cu[BLOCK_32x32].copy_cnt   = PFX(copy_cnt_32_neon);
222 -        p.cu[BLOCK_4x4].blockfill_s   = PFX(blockfill_s_4x4_neon);
223 -        p.cu[BLOCK_8x8].blockfill_s   = PFX(blockfill_s_8x8_neon);
224 -        p.cu[BLOCK_16x16].blockfill_s = PFX(blockfill_s_16x16_neon);
225 -        p.cu[BLOCK_32x32].blockfill_s = PFX(blockfill_s_32x32_neon);
226 +        p.cu[BLOCK_4x4].blockfill_s[ALIGNED]   = PFX(blockfill_s_4x4_neon);
227 +        p.cu[BLOCK_8x8].blockfill_s[ALIGNED]   = PFX(blockfill_s_8x8_neon);
228 +        p.cu[BLOCK_16x16].blockfill_s[ALIGNED] = PFX(blockfill_s_16x16_neon);
229 +        p.cu[BLOCK_32x32].blockfill_s[ALIGNED] = PFX(blockfill_s_32x32_neon);
232          p.cu[BLOCK_4x4].copy_ss   = PFX(blockcopy_ss_4x4_neon);
234          p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].copy_sp = PFX(blockcopy_sp_32x64_neon);
237 -        p.cu[BLOCK_4x4].add_ps   = PFX(pixel_add_ps_4x4_neon);
238 -        p.cu[BLOCK_8x8].add_ps   = PFX(pixel_add_ps_8x8_neon);
239 -        p.cu[BLOCK_16x16].add_ps = PFX(pixel_add_ps_16x16_neon);
240 -        p.cu[BLOCK_32x32].add_ps = PFX(pixel_add_ps_32x32_neon);
241 -        p.cu[BLOCK_64x64].add_ps = PFX(pixel_add_ps_64x64_neon);
242 +        p.cu[BLOCK_4x4].add_ps[ALIGNED]   = PFX(pixel_add_ps_4x4_neon);
243 +        p.cu[BLOCK_8x8].add_ps[ALIGNED]   = PFX(pixel_add_ps_8x8_neon);
244 +        p.cu[BLOCK_16x16].add_ps[ALIGNED] = PFX(pixel_add_ps_16x16_neon);
245 +        p.cu[BLOCK_32x32].add_ps[ALIGNED] = PFX(pixel_add_ps_32x32_neon);
246 +        p.cu[BLOCK_64x64].add_ps[ALIGNED] = PFX(pixel_add_ps_64x64_neon);
249 -        p.chroma[X265_CSP_I420].cu[BLOCK_420_4x4].add_ps   = PFX(pixel_add_ps_4x4_neon);
250 -        p.chroma[X265_CSP_I420].cu[BLOCK_420_8x8].add_ps   = PFX(pixel_add_ps_8x8_neon);
251 -        p.chroma[X265_CSP_I420].cu[BLOCK_420_16x16].add_ps = PFX(pixel_add_ps_16x16_neon);
252 -        p.chroma[X265_CSP_I420].cu[BLOCK_420_32x32].add_ps = PFX(pixel_add_ps_32x32_neon);
253 -        p.chroma[X265_CSP_I422].cu[BLOCK_422_4x8].add_ps   = PFX(pixel_add_ps_4x8_neon);
254 -        p.chroma[X265_CSP_I422].cu[BLOCK_422_8x16].add_ps  = PFX(pixel_add_ps_8x16_neon);
255 -        p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].add_ps = PFX(pixel_add_ps_16x32_neon);
256 -        p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].add_ps = PFX(pixel_add_ps_32x64_neon);
257 +        p.chroma[X265_CSP_I420].cu[BLOCK_420_4x4].add_ps[ALIGNED]   = PFX(pixel_add_ps_4x4_neon);
258 +        p.chroma[X265_CSP_I420].cu[BLOCK_420_8x8].add_ps[ALIGNED]   = PFX(pixel_add_ps_8x8_neon);
259 +        p.chroma[X265_CSP_I420].cu[BLOCK_420_16x16].add_ps[ALIGNED] = PFX(pixel_add_ps_16x16_neon);
260 +        p.chroma[X265_CSP_I420].cu[BLOCK_420_32x32].add_ps[ALIGNED] = PFX(pixel_add_ps_32x32_neon);
261 +        p.chroma[X265_CSP_I422].cu[BLOCK_422_4x8].add_ps[ALIGNED]   = PFX(pixel_add_ps_4x8_neon);
262 +        p.chroma[X265_CSP_I422].cu[BLOCK_422_8x16].add_ps[ALIGNED]  = PFX(pixel_add_ps_8x16_neon);
263 +        p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].add_ps[ALIGNED] = PFX(pixel_add_ps_16x32_neon);
264 +        p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].add_ps[ALIGNED] = PFX(pixel_add_ps_32x64_neon);
267          p.cu[BLOCK_4x4].cpy2Dto1D_shr   = PFX(cpy2Dto1D_shr_4x4_neon);
269          p.cu[BLOCK_32x32].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_32x32_neon);
272 -        p.cu[BLOCK_4x4].ssd_s   = PFX(pixel_ssd_s_4x4_neon);
273 -        p.cu[BLOCK_8x8].ssd_s   = PFX(pixel_ssd_s_8x8_neon);
274 -        p.cu[BLOCK_16x16].ssd_s = PFX(pixel_ssd_s_16x16_neon);
275 -        p.cu[BLOCK_32x32].ssd_s = PFX(pixel_ssd_s_32x32_neon);
276 +        p.cu[BLOCK_4x4].ssd_s[ALIGNED]   = PFX(pixel_ssd_s_4x4_neon);
277 +        p.cu[BLOCK_8x8].ssd_s[ALIGNED]   = PFX(pixel_ssd_s_8x8_neon);
278 +        p.cu[BLOCK_16x16].ssd_s[ALIGNED] = PFX(pixel_ssd_s_16x16_neon);
279 +        p.cu[BLOCK_32x32].ssd_s[ALIGNED] = PFX(pixel_ssd_s_32x32_neon);
282          p.cu[BLOCK_4x4].sse_ss   = PFX(pixel_sse_ss_4x4_neon);
284          p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].sub_ps = PFX(pixel_sub_ps_32x64_neon);
287 -        p.cu[BLOCK_4x4].calcresidual   = PFX(getResidual4_neon);
288 -        p.cu[BLOCK_8x8].calcresidual   = PFX(getResidual8_neon);
289 -        p.cu[BLOCK_16x16].calcresidual = PFX(getResidual16_neon);
290 -        p.cu[BLOCK_32x32].calcresidual = PFX(getResidual32_neon);
291 +        p.cu[BLOCK_4x4].calcresidual[ALIGNED]   = PFX(getResidual4_neon);
292 +        p.cu[BLOCK_8x8].calcresidual[ALIGNED]   = PFX(getResidual8_neon);
293 +        p.cu[BLOCK_16x16].calcresidual[ALIGNED] = PFX(getResidual16_neon);
294 +        p.cu[BLOCK_32x32].calcresidual[ALIGNED] = PFX(getResidual32_neon);
297          p.cu[BLOCK_4x4].sse_pp   = PFX(pixel_sse_pp_4x4_neon);